To appear in: Annual Review of Language Acquisition, Vol. 2, 2002
Language discrimination by newborns: Teasing apart phonotactic,
rhythmic, and intonational cues
Franck Ramus
Laboratoire de Sciences Cognitives et Psycholinguistique (EHESS/CNRS), Paris, France
Speech rhythm has long been claimed to be a useful bootstrapping cue in the very first steps
of language acquisition. Previous studies have suggested that newborn infants do categorize
varieties of speech rhythm, as demonstrated by their ability to discriminate between certain
languages. However, the existing evidence is not unequivocal: in previous studies, stimuli
discriminated by newborns always contained additional speech cues on top of rhythm. Here,
we conducted a series of experiments assessing discrimination between Dutch and Japanese by
newborn infants, using a speech resynthesis technique to progressively degrade non-rhythmical
properties of the sentences. When the stimuli are resynthesized using identical phonemes and
artificial intonation contours for the two languages, thereby preserving only their rhythmic
and broad phonotactic structure, newborns still seem to be able to discriminate between the
two languages, but the effect is weaker than when intonation is present. This leaves open the
possibility that the temporal correlation between intonational and rhythmic cues might actually
facilitate the processing of speech rhythm.
Key-words: newborn speech perception language discrimination rhythm intonation prosody
bootstrapping.
Language acquisition is a field notorious for its bootstrapping problems: in essence, it seems impossible to explain
how each component of language is learnt without appealing to previous knowledge of other components. How does
the child learn syntax? By relying on his/her knowledge of
words, their meaning, and the meaning of whole sentences,
as revealed by observation (this is semantic bootstrapping;
Pinker, 1984). But how does the child learn the meaning of
words? You have to assume some notions of syntax (this is
syntactic bootstrapping; Gleitman, 1990).
These apparent paradoxes have raised interest in the study
of the raw input available to the child, i.e., the speech signal,
and of how much information can be extracted thereof. In
this line, Gleitman and Wanner (1982) had already long ago
suggested that prosody (rhythm, intonation) might play an
important role in the acquisition of syntax (this was prosodic
bootstrapping). Prosody has also been shown to be an impor-
tant cue to the segmentation of continuous speech into words
(see Jusczyk, 1997 for a review). More generally, there is
of course no reason to restrict the range of potential cues to
prosody. Phonological bootstrapping, although a misnomer,
now sums up the idea that a direct surface analysis of all
sorts of acoustic and phonetic cues should be of great help to
the first steps of phonological, lexical and syntactic acquisition (Morgan & Demuth, 1996; Christophe, Guasti, Nespor,
Dupoux, & Ooyen, 1997). The present paper is dedicated to
the study of one such cue, speech rhythm.
Rhythm may be viewed as a parameter that shows variation across the languages of the world. Three types of rhythm
have been identified, leading to a classification of languages
into three classes (Pike, 1945; Abercrombie, 1967; Ladefoged, 1975): stress-timed languages, including most Germanic languages as well as Russian, Arabic or Thai, syllabletimed languages, including most Romance languages as well
as Turkish or Yoruba, and mora-timed languages, including Japanese. Historically, stress- and syllable-timing referred to the idea that stresses or syllables would be regularly paced in the corresponding languages, but this intuition
was never supported by firm empirical evidence (e.g., see
Roach, 1982; Dauer, 1983). However, more recent studies have shown that this rhythm typology can be grounded
in languages’ phonological properties (Dauer, 1987) and in
acoustic-phonetic measurements (Ramus, Nespor, & Mehler,
1999; Low, Grabe, & Nolan, 2000).
In order to know whether rhythm plays a part in language
acquisition, one needs to ask: (a) what other component of
language rhythm might help learn, (b) whether infants are
able to represent rhythmic differences, and (c) whether infants do actually use rhythm to learn that component. The
first question has already been the subject of many specu-
I thank Jacques Mehler, Marc Hauser, Anne Christophe,
Christophe Pallier, Emmanuel Dupoux and Ghislaine DehaeneLambertz for useful discussions, Sylvie Margules and Renate Zangl
for help testing the subjects, Jacques Mehler, Anne Christophe,
John Morton and Sarah Griffiths for comments on a previous version of this paper, Xavier Jeannin and Michel Dutat for technical
assistance, and the Délégation Générale pour l’Armement for financial support.
Experiments 1 and 2 have already been partially reported in Ramus, Hauser, Miller, Morris, and Mehler (2000) within a different
context, and are described here in greater detail with permission
from the American Association for the Advancement of Science.
Correspondence to: Franck Ramus, LSCP, 54 boulevard Raspail, 75006 Paris, France. E-mail: [email protected].
1
2
FRANCK RAMUS
lations. Generally speaking, if the infant could determine
very early on that her native language belongs to one of three
rhythm classes, this would divide roughly by three the size
of the search space for the correct grammar. But more concrete hypotheses have been proposed, relying on evidence
that other interesting aspects of language are congruent with
the rhythm classes.
It has been argued that adult listeners perform a pre-lexical
segmentation of the speech stream, in order to obtain representation units that are best suited to their native language:
stress units (feet?) for English, syllables for French, and
morae for Japanese (Cutler & Mehler, 1993). If this is true,
then one may want to know how the child learns which representation unit is best suited to her language. The correspondence between pre-lexical unit and rhythm (verified on those
three languages so far) suggests that sensitivity to speech
rhythm might be the answer.
Another case of such a correspondence has arisen from
studies of adaptation to fast speech. Mehler et al. (1993) have
shown that, when exposed to time-compressed sentences, listeners adapt quite quickly and soon recover asymptotic comprehension. More remarkable is the finding that exposure
to compressed speech in a foreign, unknown language, improves one’s comprehension of compressed speech in the
native language. Indeed, Spanish listeners benefit from exposure to compressed Catalan, and English listeners benefit from exposure to compressed Dutch (Pallier, SebastianGalles, Dupoux, Christophe, & Mehler, 1998). However,
cross-linguistic adaptation does not work across the board, as
shown by the failure of French listeners to benefit from exposure to compressed English. Further cross-linguistic work
suggests that transfer of adaptation to fast speech operates
between languages that belong to the same rhythm class,
but not across different rhythm classes (Sebastián-Gallés,
Dupoux, & Costa, 2000). This prompted these authors to hypothesize that speech rate normalization involves pre-lexical
processes that differ across languages, in a way that is congruent with rhythm classes. Therefore, sensitivity to speech
rhythm might enable the infant to select the correct speech
rate normalization strategy very early on, which would be of
enormous help since acoustic variability is one of the main
difficulties in the way of the acquisition of stable lexical representations.
A third parameter has been shown to vary with rhythm:
syllable structure. Indeed, stress-timed languages have complex syllables, syllable-timed languages less so, and moratimed languages a very simple syllable structure. Many
phonological phenomena suggest that syllable structure is
not just a fact about how words are in a particular language,
but a piece of abstract phonological knowledge that plays
an active role in speech production and perception (Blevins,
1995; Prince & Smolensky, 1993). To cite just one example, Japanese listeners tend to "hallucinate" vowels in
the middle of consonant clusters that are incompatible with
Japanese syllable structure (Dupoux, Kakehi, Hirose, Pallier,
& Mehler, 1999; Dupoux, Pallier, Kakehi, & Mehler, 2001).
Contrary to intuition, it is not obvious for the child to learn
the correct syllabic grammar. Indeed, compiling all the syl-
lable types would require prior syllabification of the speech
stream, which is itself dependent on the syllabic grammar.
And compiling the codas and onsets from utterances’ boundaries would not be enough, since legal codas and onsets are
often different at word boundaries and word-internally. We
have therefore proposed that rhythm might provide part of
the missing information (Ramus et al., 1999).
A link has also been proposed between rhythm and word
learning. Indeed, a large literature has documented the use
of prosodic cues for lexical access. For instance, the fact that
most English words are stressed on their first syllable leads
listeners to expect a word boundary before stresses (Cutler,
1996). Convincing evidence has been provided that this very
strategy may underlie the first steps of lexical acquisition in
English (Jusczyk, Houston, & Newsome, 1999). But other
strategies are needed in languages that have another or no
predominant stress pattern. The infant therefore first needs
to learn the correct word segmentation strategy. Again, it has
been proposed that rhythm might be the cue to the correct
strategy (Jusczyk, 1997). In this case, however, it remains
particularly unclear how good the mapping between rhythm
class and segmentation strategy is: French and Spanish are
both syllable-timed, but the former has systematic word-final
stress, and the latter a predominant penultimate stress pattern.
The various hypotheses described above are not mutually
exclusive. On the contrary, they seem to be converging on the
idea that rhythm introduces the infant to pre-lexical phonology, that is, to the representation of the speech signal that is
best suited to her native language. Rhythm may thus be a
key element of phonological bootstrapping: it won’t necessarily get the child words or syntactic constituents or categories, but it may provide a preliminary toolbox for further,
more focused analyses. In essence, it may act as a guide to
the best bootstrapping strategy. Even though the present hypotheses are only tentative at this stage, both their diversity
and convergence suggest that it is relevant to move on to the
next question, that is, whether infants are sensitive to speech
rhythm, which is the focus of the present study.
Speech rhythm perception by
infants: Available evidence
Some researchers have directly studied rhythm perception by infants. Demany, McKenzie, and Vurpillot (1977)
showed that 2-3 month-old infants are able to discriminate sequences of tones differing in temporal organization.
Fowler, Smith, and Tassinary (1986) moreover showed that
3-4 month-olds were able to discriminate sequences of syllables whose Perceptual-centers1 (Morton, Marcus, & Frankish, 1976) were isochronous or not. However, it is not clear at
all whether the notion of rhythm investigated in those studies
has anything to do with that of linguistic rhythm, as defined
above.
Another set of studies has suggested that newborns do
classify languages according to their rhythm, as revealed
1
A perceptual-center is the perceived moment of occurrence of
a syllable.
LANGUAGE DISCRIMINATION BY NEWBORNS: TEASING APART PHONOTACTIC, RHYTHMIC, AND INTONATIONAL CUES
by their capacity to discriminate between different languages (Mehler et al., 1988; Nazzi, Bertoncini, & Mehler,
1998; Ramus et al., 2000). The evidence proposed is
twofold. First, one may look at the pattern of successes
and failures accumulated across different experiments using different language pairs: newborns have been shown
to discriminate between stress-timed and syllable-timed languages (French-Russian and English-Italian: Mehler et al.,
1988; English-Spanish: Moon, Cooper, & Fifer, 1993) and
between stress-timed and mora-timed languages (EnglishJapanese: Nazzi et al., 1998; Dutch-Japanese: Ramus et al.,
2000). However, the only attempt to assess within-class
discrimination has yielded a negative result (English-Dutch:
Nazzi et al., 1998)2 . Moreover, newborns also discriminated between a set of English and Dutch (stress-timed)
sentences and a set of Spanish and Italian (syllable-timed)
sentences, but failed when the two sets (English and Italian versus Dutch and Spanish) did not reflect two types of
rhythm (Nazzi et al., 1998). Thus, they seem to be able to
discriminate between sets of languages, if and only if these
sets are congruent with different rhythm classes. As impressive as this result may be, the small number of languages
studied does not guarantee that this pattern would hold for
other unrelated languages. Indeed, considering the great variety of cues present in speech, other properties than rhythm
may have allowed discrimination, and may therefore be considered as confounding factors. This concern has led these
researchers to look for a second line of evidence, by reducing
the speech cues that were available for discrimination. Thus,
Mehler et al. (1988) successfully replicated their experiments
after low-pass filtering their stimuli at 400 Hz. This process,
which eliminates the higher frequencies of speech, and therefore most of the phonetic information, is thought to preserve
only its prosodic properties (rhythm and intonation). Similarly, the experiments by Nazzi et al. (1998) used filtered
speech exclusively, and Ramus et al. (2000) used sentences
that were resynthesized in such a way as to preserve only
prosodic cues (see below). Thus, there is converging evidence that prosody is all newborns need to discriminate languages.
Nevertheless, prosody does not reduce to rhythm. It remains possible that its other major component, intonation,
plays a role in the observed discriminations. Although we
do not know of typological studies of intonation that would
allow us to make specific predictions for all the pairs of languages considered, it is, for instance, predictable that English
and Japanese should be discriminable on the basis of their intonation. Indeed, English is a Head-Complement language,
whereas Japanese is Complement-Head (for example, relative phrases come after the corresponding verb in English,
but before it in Japanese), and this syntactic parameter is said
to have a prosodic correlate, prominence, which is signaled
both in terms of rhythm and intonation (Nespor, Guasti, &
Christophe, 1996). Moreover, there is empirical evidence
that some languages are discriminable purely by their intonation, including English and Japanese (Ramus & Mehler,
1999), English and French (Maidment, 1976, 1983) and English and Dutch (Pijper, 1983). In order to assess whether
3
newborns actually perceive linguistic rhythm, it is therefore
necessary to get rid of the intonation confound, that is to go
beyond speech filtering and remove intonation from the stimuli3 .
Ramus and Mehler (1999) have adapted a technique,
speech resynthesis, to selectively degrade the different components of speech, including rhythm and intonation. This
technique has notably been used to resynthesize different
versions of English and Japanese sentences, and assess which
components of speech were sufficient for discrimination of
the two languages. The different versions included (a) broad
phonotactics + prosody, (b) prosody, (c) rhythm only, and (d)
intonation only. Results showed that pure rhythm was sufficient for French adult subjects to discriminate between the
two languages. Pure intonation was also sufficient, but the
task seemed to be much more difficult. In the present series of experiments, we wish to apply the same rationale to
the study of language discrimination by newborns, i.e., progressively eliminate the speech cues available for discrimination, and finally assess whether linguistic rhythm is, as
hypothesized, the critical cue. Experiment 1 will assess the
discrimination of Dutch and Japanese with all cues present,
Exp. 2 the discrimination of those same sentences with only
prosodic and broad phonotactic cues, Exp. 3 with only rhythmic cues, and Exp. 4 with rhythmic and broad phonotactic
cues.
Experiment 1: Natural speech
This first experiment aims to test the discrimination of two
languages in the most unconstrained condition, using natural, unsynthesized sentences. The two languages we have
selected are Dutch and Japanese. The discrimination of this
pair of languages was previously tested in 2-3 month-old English infants, and yielded only a marginally significant result (Christophe & Morton, 1998). This was interpreted as
showing a growing focus on the native language, hence a
loss of interest in foreign ones (consistent with Mehler et al.,
1988). This pair of languages has never been tested on newborns, but it is expected to be easy to discriminate, given
the English-Japanese discrimination by French newborns obtained by Nazzi et al. (1998), and the fact that English and
Dutch are very similar in many respects, including rhythm.
Materials and Method
All the experiments included in this paper use the same
methodology unless otherwise stated. We have made various
attempts at improving certain aspects of the procedure; they
are described in detail where appropriate.
2
We also have unpublished data showing that newborns do not
discriminate between Catalan and Spanish, both syllable-timed languages.
3
The point of the present study is not, obviously, to deny that intonation can be processed by newborns and play a role in language
acquisition (see for instance Guasti, Nespor, Christophe, & Ooyen,
in press), but to ask whether rhythm is sufficient for babies to discriminate languages, and therefore whether they genuinely process
rhythm.
4
FRANCK RAMUS
Stimuli. Dutch and Japanese sentences were taken from
a corpus constituted by Nazzi (1997; Nazzi et al., 1998),
comprising short news-like sentences read by four female
native speakers per language. This corpus consists exclusively of adult-directed speech. We selected 5 sentences per
speaker, i.e., 20 sentences per language, matched in number
of syllables (15 to 19, with an average of 17) and in duration (3120 ms ±186 for Dutch, 3040 ms ±292 for Japanese,
F(1, 39) = 1.1, p = 0.3). We were also concerned about the
possibility that speakers in one language might have a higher
pitch than speakers in the other language. Average fundamental frequency4 is indeed significantly different between
the two languages: 216 Hz ±19 for Dutch, 235 Hz ±15 for
Japanese, F(1, 39) = 11.8, p = 0.001. This is compensated
for through resynthesis in Experiment 2, and we will see that
this had no influence on discrimination. Sentences in subsequent experiments were resynthesized from these 40 source
sentences, and differ only with respect to the type of synthesis that was used5 .
Experimental protocol. As is customary when testing
newborns, we used the non-nutritive sucking technique in
a habituation paradigm (Eimas, Siqueland, Jusczyk, & Vigorito, 1971). We have taken particular care to blind the experimenter with respect to the condition and to reduce direct
interventions during the test to a minimum. For this purpose,
the experiment has been programmed on a PC in such a way
as to be maximally automatized.
Experimental conditions Babies are randomly assigned to
the control or to the experimental group. In the habituation
phase, they are exposed to 10 sentences uttered by two speakers of one language. In the test phase, babies from the control group hear 10 new sentences uttered by the other two
speakers in the same language, whereas babies from the experimental group hear 10 new sentences uttered by two new
speakers in the other language. The language presented in
the habituation phase is counterbalanced across subjects, resulting in four experimental conditions. Assignment of the
subjects to the different conditions is managed by the program and withheld from the experimenter.
Procedure The test takes place in a sound-attenuated
booth, with only the experimenter and the baby inside. The
experimenter sits out of the infant’s field of vision and wears
a sound-proof headset playing masking noise. This noise
consists of four superimposed streams of all the experimental
sentences playing continuously, in order to optimally mask
the stimuli. The baby is seated in a reclining chair, and is presented with a pacifier fixed on a flexible arm. The air pressure
in the pacifier is measured by a pressure transducer, amplified
and transmitted to the computer via an analog/digital board.
The computer detects sucks and computes their relative amplitude.
During the first two minutes, the baby sucks in silence
and the computer calculates a high-amplitude (HA) threshold, such that 75% of the sucks have an amplitude above
the threshold. Subsequently, only HA sucks are considered6 .
The habituation phase then starts. Each HA suck may elicit
one sentence, but HA sucks produced while a sentence is
already playing do not trigger an additional one, and two
consecutive sentences are separated by at least 500 ms of silence. Sentences are played in a random order, directly from
the hard disk of the computer, by two loudspeakers placed
in front of the baby. The habituation phase goes on until the
habituation criterion is met: it consists in a minimum 25%
decrease in the number of HA sucks per minute for two consecutive minutes, compared with the maximum number of
sucks previously produced in 1 minute7 . When the criterion
is met, the computer switches to the test phase, which lasts
for 4 minutes. Test sentences are played in the same conditions as the habituation sentences.
Delay and rejection conditions Other factors may interfere with the test and may make it necessary to delay the shift
to the test phase or simply to discard the baby’s data. We
have tried as much as possible to have these decisions made
automatically by the computer, on the basis of the sucking
pattern and of indications given by the experimenter. When
the baby loses the pacifier, starts crying, or falls asleep, the
experimenter needs to take appropriate action. When doing
so he presses a special key on the keyboard, indicating the
occurrence of an event interfering with the baby’s sucking.
The most critical period in the test consists of the two minutes before and the two after the shift, which are used for the
statistical analyses. It is important to ensure that during those
four minutes, (a) no interference has occurred, (b) the baby
was awake and sucking, and thus heard enough sentences.
The computer thus implements the following conditions:
Delay:
• if some interference was signaled during the 2 minutes preceding the shift,
• OR if the baby didn’t hear any sentence during any
one of those 2 minutes,
then the shift is delayed for at least one minute, and the
habituation phase goes on.
Rejection:
• if some interference was signaled during the 2 minutes following the shift,
4
Fundamental frequency was extracted at intervals of 5 ms using
the Bliss software. We calculated an average F0 for each sentence,
as the average of all its non-zero F0 values.
5
Samples
of
the
different
types
of
stimuli
used in the present experiments can be heard on:
http://www.lscp.net/persons/ramus/resynth/ecoute.htm.
6
Eliminating the weaker 25% of the sucks helps increasing the
signal/noise ratio (Siqueland & DeLucia, 1969).
7
The first minute of the phase is not taken into account for the
determination of the maximum. Additional conditions impose that
this maximum is at least 20 HA sucks per minute, and that the habituation phase lasts at least 5 minutes.
5
LANGUAGE DISCRIMINATION BY NEWBORNS: TEASING APART PHONOTACTIC, RHYTHMIC, AND INTONATIONAL CUES
then the test is discontinued and the baby’s data are
discarded.
In addition, the experimenter himself may make the decision
to discontinue the test, (a) if the baby refuses the pacifier,
(b) if he/she doesn’t stay awake or suck enough, (c) if he/she
keeps crying or being agitated.
Participants. Experiments took place at the Maternité
Port-Royal, Hôpital Cochin in Paris. Participation was solicited from the mothers after birth, during their stay at the
maternity hospital. Babies were pre-selected on the basis of
their medical files, according to the following criteria:
• Age between 2 and 5 days old;
• Gestational age greater than or equal to 38 weeks;
• Birth weight greater than or equal to 2800 grams;
• No suffering at birth (APGAR score = 10 at 5 min);
• Normal medical assessments at birth and at two days;
• No seroconversion to rubella or toxoplasmosis;
• Mother not affected by viruses, and not addicted to any
drug including alcohol or nicotine;
• No family history of deafness or neurological problems;
• No Dutch or Japanese spoken at home8 .
Babies were tested three hours after feeding on average,
when they could be easily woken and kept in a quiet alert
state, and when their sucking reflex was maximal.
In the present experiment, 32 babies were successfully
tested, 18 males and 14 females, with a mean age of 67 ± 22
hours, a mean gestational age of 40 ± 1; 1 weeks and a mean
birth weight of 3530 ± 402 g. Twenty-nine came from monolingual French families, 2 from families where one or several
other languages than French are spoken and 1 from a family
where no French is spoken. The results of 42 additional babies were rejected for the following reasons: rejection of the
pacifier (1), sleeping or insufficient sucking before the shift
(12), crying or agitation (9), failure to meet the habituation
criterion (9), sleeping or insufficient sucking after the shift
(6), loss of the pacifier after the shift (4) and computer failure
(1).
Results
Figure 1 shows the number of HA sucks per minute for the
2 groups of subjects. To ensure that babies were in comparable conditions during the habituation phase, an ANOVA was
performed on average number of HA sucks over the 5 minutes preceding the shift, and showed no significant effect of
group (control or experimental) [F(1, 31) = 2.6, p = 0.12],
although there is a trend for babies in the control group to
suck more, and no significant effect of the language heard
in habituation (Dutch or Japanese) [F(1, 31) < 1]. In order to assess whether the experimental group reacted more
to the change than the control group, we conducted an analysis of covariance (ANCOVA), with the average number of
HA sucks during the 2 post-shift minutes as dependent variable, the average number of HA sucks during the 2 pre-shift
minutes as covariate, and the group as independent variable9 .
Here, there is no significant group effect [F(1, 29) < 1],
showing that the babies in the experimental group have not
reacted to the language change.
Number of HA sucks per minute
• OR if the baby didn’t hear any sentence during any
one of those 2 minutes,
• OR if the habituation phase has already lasted for 20
minutes,
60
50
40
30
20
10
0
-5
-4
-3
-2
-1
1
2
3
4
Minutes
Control Group
Experimental group
Figure 1.
Exp. 1: Dutch-Japanese discrimination – Natural
speech. Minutes are numbered from the shift, indicated by the
vertical line. Error bars represent ±1 standard error of the mean.
Adapted with permission from Ramus et al. (2000). Copyright 2000
American Association for the Advancement of Science.
Discussion
This experiment shows that newborns fail to discriminate
between Dutch and Japanese when the stimuli are not degraded at all, consisting just of natural sentences. This may
seem inconsistent with Nazzi et al.’s (1998) finding that similar French newborns can discriminate between English and
Japanese. Among the few differences between our experiment and that of Nazzi et al., is the fact that their sentences
were low-pass filtered, whereas ours aren’t. While it may
seem that the more information, the easier the discrimination, previous experiments on adults suggest that it is not
always the case: Irrelevant information may actually impair
the discrimination (Ramus & Mehler, 1999; Ramus, Dupoux,
Zangl, & Mehler, submitted).
8
Past studies show that newborns do not need to be familiar with
one of the two languages to be able to discriminate them (Mehler &
Christophe, 1995; Nazzi et al., 1998); familiarity only affects preference for one language (Mehler et al., 1988), which we are not testing here. Except for the two target languages, we therefore made no
particular effort to discard babies from other linguistic backgrounds
than French.
9
This is the standard analysis of sucking rates since Christophe,
Dupoux, Bertoncini, and Mehler (1994) showed that it is more appropriate than doing a simple ANOVA on a dishabituation index
(e.g., the difference in sucking rates between the 2 minutes after
and the 2 minutes before the shift). Here, we also ran the ANOVAs
and found that they always led to the same conclusions as the ANCOVAs. We thus only report the results of the latter.
6
FRANCK RAMUS
Here, the fact that each newborn hears 4 speakers during
the experiment may constitute such irrelevant information.
Indeed, it has been suggested that normalizing for speaker
variability is a costly process in younger infants, that may
interfere with other speech categorization abilities (Jusczyk,
Pisoni, & Mullenix, 1992). It is actually remarkable that
all language discrimination experiments performed on newborns to this day have used stimuli where speaker variability was either completely absent, through the use of a single
bilingual speaker (Mehler et al., 1988; Moon et al., 1993),
or at least strongly attenuated through the use of low-pass
filtering (Nazzi et al., 1998). Experiment 1 is thus the first
language discrimination experiment to expose newborns to 4
different voices.
Speech resynthesis is a convenient technique to test
whether newborns were disturbed by speaker variability:
whatever the original number of speakers, all sentences are
synthesized using only one voice, that of the synthesizer. If
our hypothesis is correct, we should then predict that newborns will be able to discriminate the two languages once the
sentences are resynthesized.
Experiment 2: Saltanaj
resynthesized speech
Materials and Method
Stimuli. We used the same sentences as for Experiment
1, but we resynthesized them in the saltanaj manner described by Ramus and Mehler (1999). For each sentence,
the fundamental frequency (F0 ) is measured, phonemes are
manually identified and their duration measured, with the
aid of speech analysis software. These parameters can be
manipulated at will, and then used as input to the speech
synthesizer MBROLA10 (Dutoit, Pagel, Pierret, Bataille, &
Vrecken, 1996). The saltanaj manipulation involves reducing the phonemic inventory of the sentences, by mapping
each phoneme to a single instance of the same manner of
articulation: fricatives are mapped to /s/, vowels to /a/, liquids to /l/, plosives to /t/, nasals to /n/ and glides to /j/. The
exact phoneme durations as well the F0 curve are copied
from the original sentences. Thus, the overall rhythm and
intonation of the sentences are faithfully preserved, together
with broad phonotactic characteristics. However, phonetic
and fine phonotactic differences are eliminated. Obviously,
access to syntax and meaning is blocked as well. Voice
differences are eliminated, since a single synthetic voice is
used; however, prosodic differences between speakers are
preserved. It therefore still makes sense to have a speaker
change in the control condition, where "speakers" refers to
those who produced the original sentences.
Incidentally, resynthesis gives full control over the average fundamental frequency which, we have noted earlier, was
significantly different between the two languages. We have
thus decided to reduce this difference by multiplying all F0
values by 1.04 for Dutch, and by 0.96 for Japanese. Note that
although this manipulation removes a possible confound, it
is not a very plausible one. Indeed, in Experiment 1, the fact
that Japanese speakers have a higher pitch on average isn’t
sufficient by itself to allow for a discrimination.
Procedure. The procedure was the same as for Exp. 1,
except that we tried to adapt the 2-shift design of Hesketh,
Christophe, and Dehaene-Lambertz (1997) to experimentation on newborns. The first two phases of the experiment
were run exactly as in Exp. 1. After a baby had undergone
the first shift and the four-minute test phase, a second shift
was made possible. It was subject to the same habituation criterion, delay and rejection conditions as the first one. When
met, it allowed the baby to undergo a second four-minute test
phase. In that case, the second shift was of a different kind
(language or speaker) to the first one. Babies who didn’t
successfully pass the second shift were nevertheless kept for
analysis of the first shift. Thus, this attempt at improving
the procedure did not interfere with the collection of the data
on the first shift, making it possible to independently analyze
the results of the two shifts. Indeed, for all babies, everything
was as in Exp. 1 up until the fourth test minute. Some babies
were just allowed to go on further if they could. This experiment was stopped as soon as 32 babies successfully passed
the first shift.
Participants. Thirty-two babies were successfully tested,
16 males and 16 females, with a mean age of 80 ± 25 hours,
a mean gestational age of 39 ± 0; 6 weeks and a mean birth
weight of 3508 ± 477 g. Twenty-five came from monolingual French families, 3 from families where one or several
other languages than French are spoken and 4 from a family where no French is spoken. The results of 20 additional
babies were rejected for the following reasons: sleeping or
insufficient sucking before the first shift (6), crying or agitation (4), failure to meet the habituation first criterion (1),
sleeping or insufficient sucking after the first shift (3), loss of
the pacifier after the first shift (6).
Results
Figure 2 shows the number of HA sucks per minute for
the two groups of babies. There was no significant group
effect on babies’ sucking during the 5 pre-shift minutes
[F(1, 31) < 1], neither was there an effect of the habituation language [F(1, 31) = 1.8, p = 0.19]. An ANCOVA
on the 2 post-shift minutes, controlling for the 2 pre-shift
minutes, showed a significant group effect [F(1, 29) = 6.3,
p = 0.018], indicating that babies in the experimental group
increased their sucking significantly more than those in the
control group. This leads us to the conclusion that the babies
in the experimental group were able to discriminate between
the two languages.11
10
MBROLA
is
freely
available
from
http://tcts.fpms.ac.be/synthesis/mbrola.html.
11
Out of 32 babies successfully passing the first shift, only 11
passed the second one. The results of the remaining 21 were discarded due to sleeping or insufficient sucking before the second
shift (7), crying or agitation (2), failure to meet the second habituation criterion before the 20 minute time limit (6), sleeping or in-
LANGUAGE DISCRIMINATION BY NEWBORNS: TEASING APART PHONOTACTIC, RHYTHMIC, AND INTONATIONAL CUES
Number of HA sucks per minute
60
50
40
30
20
10
0
-5
-4
-3
-2
-1
1
2
3
4
Minutes
Control group
Experimental group
Figure 2. Exp. 2: Dutch-Japanese discrimination – Saltanaj
speech, first shift. Minutes are numbered from the shift, indicated
by the vertical line. Error bars represent ±1 standard error of the
mean. Adapted with permission from Ramus et al. (2000). Copyright 2000 American Association for the Advancement of Science.
Discussion
The data obtained on the 32 newborns who successfully
passed the first shift show that (a) they are able to discriminate between Dutch and Japanese, (b) they can do so when
sentences are resynthesized in the saltanaj manner, i.e. when
lexical, syntactic, phonetic and some phonotactic information is removed.
Although the interaction with Exp. 1 is not quite significant [F(1, 59) = 2.6, p = 0.11]12 , this is also consistent with
the hypothesis that newborns have difficulties coping with
talker variability (Jusczyk et al., 1992), which would be the
reason why they failed to discriminate the same sentences
when they were not resynthesized.
The saltanaj resynthesis achieves a comparable level of
stimulus degradation as low-pass filtering: Since all the durations and the fundamental frequency are faithfully reproduced, prosody, in a broad sense, is still preserved. It is
therefore insufficient to disentangle the role of rhythm and
intonation. This concern is addressed in the next two experiments.
Experiment 3: sasasa with
artificial intonation
Materials and Method
Stimuli. In previous experiments testing language discrimination by adults on the basis of rhythm only (Ramus
& Mehler, 1999; Ramus et al., submitted), sentences were
resynthesized in the flat sasasa manner: all consonants were
mapped to /s/ and all vowels to /a/, and in addition the original F0 contour of the sentence was ignored and replaced by
a constant F0 . Thus all differences concerning intonation or
syllable structure were eliminated, preserving only rhythmical differences between the two languages.
7
When testing babies, an additional concern is to keep them
awake and active in the experiment. In this respect, flat
sasasa stimuli are potentially problematic. Both their low
phonetic diversity and their monotonous intonation are susceptible to provoke boredom or distress in infants, and/or
to induce them to process the stimuli as non-speech. We
thus felt we had to improve the attractiveness of our stimuli,
while still adequately testing our hypotheses. Considering
that newborns are known to react normally to low-pass filtered speech (Mehler et al., 1988; Nazzi et al., 1998), we assumed that phonetic diversity was not a necessary condition,
but we chose to preserve some variability in the intonation.
Therefore, we decided to resynthesize the same 40 sentences as before using a sasasa phonetic mapping, i.e., to
map each consonant to /s/ and each vowel as /a/. However,
instead of applying a flat intonation to each sentence, we applied artificial intonation contours. Five intonation contours
inspired from French sentences were designed and each was
applied to 4 of the Dutch and 4 of the Japanese sentences.
All contours included a regular declination towards their end,
in order to be more easily adapted to sentences of different
lengths; they are illustrated in Figure 3. Thus, the resynthesized sentences incorporate both within-sentence and withinlanguage intonational variability, but no differences between
the two languages.
A potential criticism of this method is that there might be
an interaction between intonation and rhythmic structure, so
that the five contours selected might be more adapted to the
rhythmic structure, say, of Dutch, than to that of Japanese.
This would then introduce a difference between the two sets
of sentences that would not be a rhythmic difference, strictly
speaking. In order to investigate this possibility, we conducted the following preliminary experiment on adult subjects.
Judgement by adult subjects of sentences resynthesized
with an artificial intonation Twelve participants were recruited and tested in a quiet room. They were 4 men and 8
women, with a mean age of 34 years, and of various native
languages (4 French, 1 Rumanian, 3 Spanish, 1 German, 2
English, 1 Dutch).
Two blocks of stimuli were designed: one comprising the
40 experimental sentences to be used on the babies (sasasa
sufficient sucking after the second shift (4), loss of the pacifier after
the second shift (2). The small proportion of babies able to undergo
the second shift already shows that this procedure, as used on 2month-olds, is not viable for newborns: it would lead to discard
too much data (here, a total of 41 babies out of 52). In addition, a
discrimination index computed as in Hesketh et al., 1997 was not
significantly different from 0 [t(10) < 1], indicating that these 11
babies did not increase their sucking more after the experimental
shift than after the control one. Rather, they tended to persevere
in the behavior elicited by the first shift. We conclude that there is
little to learn from a second shift with newborns.
12
Note that interactions are seldom significant in experiments on
newborns anyway, due to their low statistical power. For instance,
in directly comparable studies, no significant interaction were ever
reported (Mehler et al., 1988; Nazzi et al., 1998).
8
FRANCK RAMUS
Contour 15
400
Fundamental
frequency (Hz)
with artificial intonation), and the other the saltanaj sentences with original intonation used in Exp. 2, to provide
a baseline.
Each participant heard the sentences one by one, in a random order within each block; the order of the blocks was
counterbalanced across subjects. The task was to judge how
natural the intonation of each sentence was (from 0: very
strange to 5: perfectly natural). If artificial intonations are
equally adapted to the rhythmic structure of the two languages, they should yield similar average judgements. These
are shown in Figure 4.
300
200
100
0
0
1000
2000
3000
Dutch
Fundamental
frequency (Hz)
Average intonation ranking
Contour 16
400
300
200
100
0
0
1000
2000
3000
Contour 17
Fundamental
frequency (Hz)
400
300
200
3
2.5
2
saltanaj
sasasa - artificial
intonation
Synthesis type
0
1000
2000
3000
Contour 18
400
Fundamental
frequency (Hz)
3.5
100
0
300
200
100
0
0
1000
2000
3000
2000
3000
Contour 19
400
Fundamental
frequency (Hz)
Japanese
300
200
100
0
0
1000
Tim e (m s)
Figure 3. Intonation contours used in Exp. 3.
Figure 4. Adult subjects’ judgements of the intonation pattern of
Dutch and Japanese sentences. Error bars represent ±1 standard
error of the mean by subject.
It appears that the artificial intonations of the sasasa stimuli are judged to be significantly less natural in Dutch than in
Japanese sentences [F(1, 11) = 12.5, p = 0.005]. However,
the same is true of the saltanaj sentences with their original
intonation [F(1, 11) = 8.4, p = 0.02]. It thus can’t be interpreted as an effect of mismatch between the artificial intonation contours and Dutch rhythm. Rather, it seems to reflect
the influence of syllabic structure on subjects’ responses, although they were instructed to report specifically about intonation. From their reports, it appears that the presence of
heavy consonant clusters in Dutch (also reflected by longer
/s/s in the sasasa version) biased them in favor of Japanese.
Thus, there is a main effect of language [F(1, 11) = 21.8,
p = 0.001], and there is also a main effect of type of synthesis [F(1, 11) = 8.4, p = 0.02], revealing that sasasa synthesis
sounded less natural to the subjects than saltanaj. Despite
these effects that interfered with the subjects’ judgements,
it is most important to note that there is no interaction between language and type of stimuli [F(1, 11) < 1], indicating
that the artificiality of the sasasa stimuli’s intonation did not
interact with the rhythmic structure of the two languages13 ,
which was the hypothesis under test. We therefore consider
13
Note also that no floor nor ceiling effect would have prevented
this interaction to emerge.
9
LANGUAGE DISCRIMINATION BY NEWBORNS: TEASING APART PHONOTACTIC, RHYTHMIC, AND INTONATIONAL CUES
Procedure. Due to the unsuccessful attempt to have the
babies undergo 2 shifts in Exp. 2, we abandoned the second
shift and returned to the one-shift procedure used in Exp. 1.
Participants. Forty babies14 were successfully tested, 19
males and 21 females, with a mean age of 66 ± 23 hours, a
mean gestational age of 40; 1 ± 1; 1 weeks and a mean birth
weight of 3428 ± 424 g. Twenty-six came from monolingual
French families, 11 from families where one or several other
languages than French are spoken and 3 from families where
no French is spoken. The results of 20 additional babies were
rejected for the following reasons: rejection of the pacifier
(2), sleeping or insufficient sucking before the shift (7), crying or agitation (3), failure to meet the habituation criterion
(1), sleeping or insufficient sucking after the shift (4), loss of
the pacifier after the shift (3).
60
Number of HA sucks per minute
our stimuli as appropriate to test language discrimination on
the basis of rhythm only.
50
40
30
20
10
0
-5
-4
-3
-2
-1
1
2
3
4
Minutes
Control group
Experimental group
Figure 5. Exp. 3: Dutch-Japanese discrimination – Sasasa speech
with artificial intonation. Minutes are numbered from the shift, indicated by the vertical line. Error bars represent ±1 standard error
of the mean.
Results
Figure 5 shows the number of HA sucks per minute for the
two groups of babies. There was no significant group effect
on babies’ sucking during the 5 pre-shift minutes [F(1, 39) =
1.5, p = 0.23]. However, there was a significant effect of the
habituation language [F(1, 39) = 12.8, p = 0.001], babies
listening to Dutch sucking more (35.6 ± 8.3 sucks per minute
on average) than those listening to Japanese (25.8±9.1 sucks
per minute). To take this effect into account, we included the
habituation language factor in the usual ANCOVA. We found
that it had no significant effect on sucking patterns during the
2 post-shift minutes [F(1, 35) < 1], and most importantly,
that it did not interact with the group factor [F(1, 35) < 1].
Thus, this effect had no consequence on the overall result of
the experiment; its interpretation will be addressed in a later
section. Finally, the ANCOVA on the 2 post-shift minutes,
controlling for the 2 pre-shift minutes, showed no group effect [F(1, 35) < 1], indicating that the babies didn’t discriminate between the two languages.
Discussion
It is worth noting that, although the sucking patterns
around the shift of language unambiguously show an absence
of dishabituation to the new language, the fact that babies
sucked significantly more to listen to Dutch than to Japanese
in the habituation phase yet suggests that the two languages
are not entirely the same to them. This might, of course, be
a sampling effect, i.e., babies who intrinsically suck more
being assigned by chance to the "Dutch first" condition; yet,
both the size and the significance of the effect make this interpretation unlikely. We now leave this issue aside to return
to it in a later section.
There are several possible interpretations of the failure of
babies to discriminate between Dutch and Japanese given
the sasasa sentences with artificial intonation. One is that
babies don’t process speech rhythm, and that language discrimination experiments should be re-interpreted as revealing the processing of intonation. Another is that rhythm and
intonation are inseparable: babies may be sensitive to speech
rhythm, but intonation is necessary to fully process it. Yet another interpretation would be that babies can process speech
rhythm, but the stimuli used were inadequate for them to exhibit this ability.
For instance, Ramus et al. (2000) showed that newborns
don’t discriminate Dutch from Japanese anymore when the
same sentences are played backwards. This suggests that
stimuli that are not enough speech-like are not correctly processed by babies, even though they contain enough basic
acoustic information for the discrimination to be feasible in
principle. In this respect, sasasa might not be speech-like
enough: there is indeed no natural language with so little
phonetic diversity. It could also be that these stimuli are too
boring or distressing for the babies, as we had hypothesized
regarding a flat intonation. Whatever the appropriate explanation, we will now try to increase the chances that the babies
correctly process the stimuli.
Experiment 4: saltanaj with
artificial intonation
Materials and Method
Stimuli. There are two differences between the stimuli
used in Experiment 2 and those of Experiment 3: one is the
reduction of the phonetic inventory (from saltanaj to sasasa),
and the other is the use of artificial intonation contours instead of the original ones. At least one of them has caused
babies to fail in the discrimination task. It is therefore natural to undo one of those changes in order to know which was
critical. We thus returned to the saltanaj stimuli of Exp. 2,
but this time we applied them the artificial intonation contours of Exp. 3.
14
Eight additional babies were tested after a first analysis on the
first 32 babies, because at that stage it was not clear if there was a
trend that was not significant by lack of power, or no effect at all.
10
FRANCK RAMUS
Participants. Forty babies were successfully tested, 21
males and 19 females, with a mean age of 68 ± 21 hours,
a mean gestational age of 40; 1 ± 1 weeks and a mean birth
weight of 3512 ± 341 g. Twenty-seven came from monolingual French families, 12 from families where one or several
other languages than French are spoken and one from a family where no French is spoken. The results of 44 additional
babies were rejected for the following reasons: rejection of
the pacifier (5), sleeping or insufficient sucking before the
shift (22), crying or agitation (8), failure to meet the habituation criterion (2), sleeping or insufficient sucking after the
shift (4), loss of the pacifier after the shift (3).
Results
Figure 6 shows the number of HA sucks per minute for
the two groups of babies. There was no significant group
effect on babies’ sucking during the 5 pre-shift minutes
[F(1, 39) < 1], neither was there an effect of the habituation language [F(1, 39) = 2.1, p = 0.16]. An ANCOVA
on the 2 post-shift minutes, controlling for the 2 pre-shift
minutes, shows no significant group effect [F(1, 37) = 1.46,
p = 0.24]. However, examination of Figure 6 suggests that
there is an effect, which is confined to the first minute after
the shift. A new ANCOVA, taking as dependent variable the
number of sucks during the first post-shift minute, and controlling for the 2 pre-shift minutes, yields a significant group
effect indeed [F(1, 37) = 4.48, p = 0.04]. This suggests that
the newborns have again discriminated between Dutch and
Japanese. However, the effect is weaker than in Experiment
2, being evident during only one minute following the language change.
Number of HA sucks per minute
50
40
30
20
10
0
-5
-4
-3
-2
-1
1
2
3
4
Minutes
Control group
Experimental group
Figure 6. Exp. 4: Dutch-Japanese discrimination – Saltanaj
speech with artificial intonation. Minutes are numbered from the
shift, indicated by the vertical line. Error bars represent ±1 standard
error of the mean.
Discussion
Obviously, this last experiment would need to be replicated. From a methodological point of view, it is not very
satisfying to first look where the effect is located (here, on the
first minute post-shift), and then restrict the analysis accordingly, since this increases the risk of a Type I error. However,
although analyses based on two minutes before and two minutes after the shift have emerged as the methodological standards in the literature, deviations from this norm are by no
means exceptional: analyses restricted to the first post-shift
minute (e.g., McAdams & Bertoncini, 1997) or extended to
the three post-shift minutes (e.g., Nazzi et al., 1998, Exp. 3)
sometimes appear necessary to establish an effect. Awaiting
replication, we will assume that the present discrimination is
reliable for the remainder of the discussion.
Since no intonational difference remained between the
two languages in the present saltanaj stimuli, intonation is
not likely to be the cue whose processing is responsible for
the language discrimination patterns observed. Here, the
only cues available for discrimination were the sentences’
rhythm and their broad phonotactic patterns. This leaves us
with two possible interpretations.
Considering that newborns did not discriminate Dutch
from Japanese in Exp. 3, when only rhythm was available,
but did discriminate when some phonotactic information was
added to rhythm in Exp. 4, the most straightforward interpretation is that newborns actually discriminated between the
respective phonotactic patterns of Dutch and Japanese, e.g.,
they noticed the fact that there are many consonant clusters
in Dutch but not in Japanese. However, such an interpretation is at odds with quite a large body of evidence. Indeed,
sensitivity to phonotactic differences has been directly tested
in experiments where newborns had to discriminate between
lists of words of different syllabic structure. For instance,
newborns were unable to discriminate bi-syllabic words with
complex syllabic structure (e.g., CVCCCV, CCVCCV, CVCCVC...) from bi-syllabic words with simple syllabic structure (e.g., CVCV, VCCV, VCVC...), although they were
able to discriminate simple bi-syllabic (CVCV) from trisyllabic (CVCVCV) words (Bijeljac-Babic, Bertoncini, &
Mehler, 1993). Similarly, they were unable to discriminate between tri-moraic (CVCCV) and bi-moraic words
(CVCV), although they were again able to discriminate bisyllabic from tri-syllabic words (Bertoncini, Floccia, Nazzi,
& Mehler, 1995). If newborns were able to extract phonotactic regularities from 3-second long Dutch and Japanese utterances, they would be expected to do so also from bi-syllabic
words. The fact that they do not suggests that sensitivity to
phonotactic differences is not available at birth; this is also
consistent with evidence that familiarity with the native language’s phonotactic pattern emerges between 6 and 9 months
of age (Jusczyk, Cutler, & Redanz, 1993; Friederici & Wessels, 1993; Jusczyk, Luce, & Charles-Luce, 1994), while familiarity with the native language’s prosodic pattern seems
to emerge soon after birth (Mehler et al., 1988; Christophe &
Morton, 1998).
The other possible interpretation of our results is that newborns perceived the rhythmical differences between the two
languages. This interpretation assumes that the reason for
the failure to discriminate in Exp. 3 may lie in some aspect
of the sasasa stimuli that prevented the babies from correctly
processing them. For instance, their extremely low phonetic
LANGUAGE DISCRIMINATION BY NEWBORNS: TEASING APART PHONOTACTIC, RHYTHMIC, AND INTONATIONAL CUES
11
ference is significant: F(1, 142) = 11.1, p = 0.001. The fact
that this effect is consistent across 4 experiments suggest that
this is not a random sampling effect. It is remarkable that a
similar effect was also noted by Nazzi et al. (1998): they
found that French newborns sucked more to listen to English than to listen to Japanese. In earlier studies, such patterns had been interpreted as showing babies’ "preference"
for or at least familiarity with their native-language: Mehler
et al. (1988) found that French newborns sucked more during the habituation phase to listen to French than to Russian,
and Moon et al. (1993), using a technique directly assessing preference, found that newborns sucked more to listen
to their native language (English or Spanish). Similar results have been found using preferential looking procedures
in older babies (Dehaene-Lambertz & Houston, 1998; Bosch
& Sebastián-Gallés, 1997).
Post-hoc analysis: Increased
Here (and also in Nazzi et al.’s study), where neither lansucking for Dutch
guage was the babies’ native language, one may want to
look for an alternative explanation. For instance, followAfter finding a significant effect of habituation language
ing adult subjects’ judgements in Exp. 3, one might argue
in Experiment 3 (with babies sucking more to Dutch than to
that Dutch sentences, containing more consonant clusters
Japanese), we have looked for a similar trend in the other exand more frication, are less pleasant to listen to, and thus
periments. As shown in Table 1 and Figure 7, the same trend
keep the babies more excited. Such an interpretation would,
was present in all four experiments, suggesting a consistent
however, assume that stimuli that are less preferred provoke
phenomenon.
more sucking, which would appear to be in contradiction
with the earlier studies’ interpretations. In the absence of
Table 1
a good model of what provokes a baby’s sucks, the question
Average number of HA sucks produced during the 5 pre-shift
remains open15 .
minutes, as a function of Experiment and language.
It is yet possible to provide an interpretation of the present
Exp. 1
Exp. 2
Exp. 3
Exp. 4 results in terms of genuine preference or familiarity. InDutch
33 ± 11
32.2 ± 11.2 35.6 ± 8.3 29.4 ± 9.6 deed, French can be seen as closer to Dutch and English
Japanese 30.9 ± 11.1 27 ± 10.4
25.8 ± 9.1 25.3 ± 8.3 than to Japanese along a number of dimensions. Regarding
the most relevant one, rhythm, objective acoustic/phonetic
measures of rhythmic properties suggest that French rhythm
is much closer to that of Dutch and English than to that of
Japanese (Ramus et al., 1999). Similar arguments could be
50
made for syllable structure, intonation, size of the phonemic
inventory... It is thus conceivable that Dutch and English
40
sound more familiar to the French newborn than Japanese16 .
Native-language preference might therefore be re-interpreted
30
as preference for the most familiar stimulus (along the dimensions that are relevant to the baby).
20
Obviously, these interpretations are very tentative and
should not be taken as strong claims. If one really wants to
10
test whether French newborns have a preference for Dutch
or English versus Japanese, then one should use a proce0
dure specially designed to assess preference, not discrimi-5
-4
-3
-2
-1
nation. The fact remains that the trend is present in every
Minutes
single of five experiments to date, and awaits an explanaDutch
Japanese
tion. This also suggests that when this effect reaches signifiNumber of HA sucks per minute
diversity might lead babies to process them as non-speech.
Alternatively, the constant frication of sasasa sentences may
have been too distressing for babies to correctly perform the
task; indeed, adult subjects also rated these stimuli lower
than the saltanaj in Exp. 3, and complained about the harsh
sound of the /s/s. Whatever the correct explanation may be,
the alternatives could be tested by running further discrimination experiments while manipulating the nature and the
variety of the phonemes used in the resynthesis. Possibly
mamama stimuli would be preferable to sasasa. One could
also resynthesize the sentences as salatanaja, a transformation similar to the sasasa in that consonant clusters would
be mapped to a single phoneme of the same total duration,
but the nature of this phoneme would be allowed to vary randomly. This will be matter for future investigations.
Figure 7. Average number of HA sucks produced during each of
the 5 pre-shift minutes, as a function of language. Data of Experiments 1-4 are collapsed. Error bars represent ±1 standard error of
the mean.
Overall, during the 5 minutes preceding the shift, babies
produced on average 32.5 ± 10 HA sucks per minute to listen to Dutch, and 27.1 ± 9.7 to listen to Japanese. The dif-
15
Note that this problem is not particular to the sucking behavior. In studies using preferential listening techniques, preference
sometimes goes for the novel stimulus, and sometimes for the familiar one. No generalized account of infants’ preferences has been
proposed.
16
Although newborns’ native language was not strictly controlled
in the present experiments, a large majority of babies had French as
the main language in their family, and none had Dutch or Japanese.
12
FRANCK RAMUS
cance (p = 0.001) in Exp. 3, it can hardly be dismissed as a
sampling effect. However, we will not go as far as to claim
that this differential processing of the two languages in the
habituation phase amounts to discrimination, when a direct
measure of discrimination provides no evidence thereof. At
this point, we have to admit that no framework is available to
interpret such effects.
Conclusion
The literature on infant speech perception has suggested
for years that newborns discriminate languages’ rhythmic
patterns. The evidence accumulated so far, although compelling, has always left open the possibility that the discriminations observed might be due to a sensitivity to intonational
differences. Here, using resynthesized stimuli, we have dissociated intonation from rhythm. The results suggest that
intonation was not necessary for newborns to discriminate
between Dutch and Japanese (Exp. 4), and therefore that
rhythmic differences between these two languages have been
processed by newborns. However, this interpretation is limited in two ways.
First, newborns appeared to discriminate the two languages only when some phonotactic information was present
(saltanaj, Exp. 4), not when purely rhythmic cues were isolated (sasasa, Exp. 3). As we discussed earlier, there is relatively convincing previous evidence that sensitivity to phonotactic differences is not available at such an early age. The
interpretation we favor therefore appeals to the inappropriateness of the sasasa stimuli for babies. Of course, this hypothesis would itself need to be bolstered by further experiments using more appropriate, yet purely rhythmic, stimuli.
This will be matter for future research.
Another problem is that, even with the phonotactic cues
present in Exp. 4, the discrimination effect found was relatively weak. This obviously invites an attempt to replicate.
We may be luckier and find a stronger effect than the one
presented here. Alternatively, we might also confirm that
pure rhythmic differences do not elicit such reliable discriminations as differences along several dimensions. It is indeed
possible that, when dissociated from intonation, rhythm loses
part of its salience. After all, in real speech, rhythm and
intonation are highly correlated: for instance, stress is signalled in terms of both the duration and the pitch of the syllable (and its energy); similarly, phonological phrase boundaries are marked both by final syllable lengthening and pitch
movements. Thus, intonation partly underlines the information provided by purely durational cues. This means that,
in addition to providing specific intonational cues, it introduces redundant rhythmic cues that may help rhythm processing itself. Indeed, it has been suggested that integrating multiple correlated cues might be a more powerful learning strategy than just considering isolated cues (Shi, Morgan, & Allopenna, 1998; Christiansen, Allen, & Seidenberg,
1998). Therefore, the most conservative interpretation of the
present set of results at this stage may be that the integration
of intonational and rhythmic cues better serves language discrimination than either of these cues taken separately. One
could further test this interpretation by presenting resynthesised sentences where intonation would be preserved, but
rhythm would be made the same for the two languages: one
would predict a similar pattern of results as in Exp. 4, i.e.,
a weaker effect than in Exp. 2. Another prediction would
be that when rhythmic and intonational differences are both
present, but de-correlated (e.g., by permuting intonation contours across sentences within each language), language discrimination would be similarly impaired, as compared to the
condition where the cues are correlated.
In conclusion, a definitive interpretation of the present set
of experiments will have to await future results. Rhythm is
still a good candidate as a perceptual cornerstone for the
baby. But it may be futile to try to isolate the effects of
purely rhythmic regularities, as their integration with other
correlated cues may be the key to a successful processing of
the speech signal.
References
Abercrombie, D. (1967). Elements of general phonetics. Chicago:
Aldine.
Bertoncini, J., Floccia, C., Nazzi, T., & Mehler, J. (1995). Morae
and syllables: Rhythmical basis of speech representations in
neonates. Language and Speech, 38, 311-329.
Bijeljac-Babic, R., Bertoncini, J., & Mehler, J. (1993). How do
four-day-old infants categorize multisyllabic utterances? Developmental Psychology, 29, 711-721.
Blevins, J. (1995). The syllable in phonological theory. In J. A.
Goldsmith (Ed.), The handbook of phonological theory (p. 206244). Cambridge: Blackwell.
Bosch, L., & Sebastián-Gallés, N. (1997). Native language recognition abilities in 4-month-old infants from monolingual and bilingual environments. Cognition, 65, 33-69.
Christiansen, M., Allen, J., & Seidenberg, M. (1998). Learning
to segment speech using multiple cues: A connectionist model.
Language and Cognitive Processes, 13, 221-268.
Christophe, A., Dupoux, E., Bertoncini, J., & Mehler, J. (1994).
Do infants perceive word boundaries? An empirical study of the
bootstrapping of lexical acquisition. Journal of the Acoustical
Society of America, 95(3), 1570-1580.
Christophe, A., Guasti, T., Nespor, M., Dupoux, E., & Ooyen, B.
van. (1997). Reflections on phonological bootstrapping: Its role
for lexical and syntactic acquisition. Language and Cognitive
Processes, 12(5/6), 585-612.
Christophe, A., & Morton, J. (1998). Is Dutch native English? Linguistic analysis by 2-month-olds. Developmental Science, 1(2),
215-219.
Cutler, A. (1996). Prosody and the word boundary problem. In
J. L. Morgan & K. Demuth (Eds.), Signal to Syntax: Bootstrapping from Speech to Grammar in Early Acquisition (p. 87-99).
Mahwah, NJ: Lawrence Erlbaum Associates.
Cutler, A., & Mehler, J. (1993). The periodicity bias. Journal of
Phonetics, 21, 103-108.
Dauer, R. M. (1983). Stress-timing and syllable-timing reanalyzed.
Journal of Phonetics, 11, 51-62.
Dauer, R. M. (1987). Phonetic and phonological components of
language rhythm. In XIth International Congress of Phonetic
Sciences (Vol. 5, p. 447-450). Tallinn.
LANGUAGE DISCRIMINATION BY NEWBORNS: TEASING APART PHONOTACTIC, RHYTHMIC, AND INTONATIONAL CUES
Dehaene-Lambertz, G., & Houston, D. (1998). Faster orientation
latencies toward native language in two-month old infants. Language and Speech, 41(1), 21-43.
Demany, L., McKenzie, B., & Vurpillot, E. (1977). Rhythm perception in early infancy. Nature, 266(5604), 718-9.
Dupoux, E., Kakehi, K., Hirose, Y., Pallier, C., & Mehler, J. (1999).
Epenthetic vowels in Japanese: A perceptual illusion? Journal of Experimental Psychology: Human Perception and Performance, 25(6), 1568-1578.
Dupoux, E., Pallier, C., Kakehi, K., & Mehler, J. (2001). New
evidence for prelexical phonological processing in word recognition. Language and Cognitive Processes, 16(5/6), 491-505.
Dutoit, T., Pagel, V., Pierret, N., Bataille, F., & Vrecken, O. van der.
(1996). The MBROLA Project: Towards a set of high-quality
speech synthesizers free of use for non-commercial purposes. In
ICSLP’96. Philadelphia. (MBROLA is freely available from
http://tcts.fpms.ac.be/synthesis/mbrola.html)
Eimas, P. D., Siqueland, E. R., Jusczyk, P. W., & Vigorito, J. (1971).
Speech perception in infants. Science, 171, 303-306.
Fowler, C., Smith, M., & Tassinary, L. (1986). Perception of syllable timing by prebabbling infants. Journal of the Acoustical
Society of America, 79(3), 814-825.
Friederici, A., & Wessels, J. (1993). Phonotactic knowledge of
word boundaries and its use in infant speech perception. Perception & Psychophysics, 54(3), 287-295.
Gleitman, L. (1990). The structural sources of verb meanings.
Language Acquisition, 1, 3-55.
Gleitman, L., & Wanner, E. (1982). The state of the state of the art.
In E. Wanner & L. Gleitman (Eds.), Language acquisition: The
state of the art (p. 3-48). Cambridge UK: Cambridge University
Press.
Guasti, M. T., Nespor, M., Christophe, A., & Ooyen, B. van. (in
press). Pre-lexical setting of the head-complement parameter
through prosody. In J. Weissenborn & B. Höhle (Eds.), How
to get into language: Approaches to bootstrapping in early language development. Amsterdam: Benjamins.
Hesketh, S., Christophe, A., & Dehaene-Lambertz, G. (1997). Nonnutritive sucking and sentence processing. Infant Behavior and
Development, 20, 263-269.
Jusczyk, P. W. (1997). The discovery of spoken language. Cambridge, MA: MIT Press.
Jusczyk, P. W., Cutler, A., & Redanz, N. (1993). Infants’ preference for the predominant stress patterns of English words. Child
Development, 64, 675-687.
Jusczyk, P. W., Houston, D. M., & Newsome, M. (1999). The beginnings of word segmentation in English-learning infants. Cognitive Psychology, 39(3/4), 159-207.
Jusczyk, P. W., Luce, P. A., & Charles-Luce, J. (1994). Infants’ sensitivity to phonotactic patterns in the native language. Journal
of Memory and Language, 33, 630-645.
Jusczyk, P. W., Pisoni, D. B., & Mullenix, J. (1992). Some consequences of stimulus variability on speech processing by 2month-old infants. Cognition, 43, 253-291.
Ladefoged, P. (1975). A course in phonetics. New York: Harcourt
Brace Jovanovich.
Low, E. L., Grabe, E., & Nolan, F. (2000). Quantitative characterisations of speech rhythm: Syllable-timing in Singapore English.
Language and Speech, 43(4), 377-401.
Maidment, J. A. (1976). Voice fundamental frequency characteristics as language differentiators. Speech and hearing: Work in
progress, University College London, 74-93.
13
Maidment, J. A. (1983). Language recognition and prosody: further evidence. Speech, hearing and language: Work in progress,
University College London, 1, 133-141.
McAdams, S., & Bertoncini, J. (1997). Organization and discrimination of repeating sound sequences by newborn infants. Journal Of The Acoustical Society Of America, 102(5), 2945-53.
Mehler, J., & Christophe, A. (1995). Maturation and learning of
language during the first year of life. In M. S. Gazzaniga (Ed.),
The Cognitive Neurosciences (p. 943-954). Bradford Books /
MIT Press.
Mehler, J., Jusczyk, P., Lambertz, G., Halsted, N., Bertoncini, J., &
Amiel-Tison, C. (1988). A precursor of language acquisition in
young infants. Cognition, 29, 143-178.
Mehler, J., Sebastián-Gallés, N., Altmann, G., Dupoux, E.,
Christophe, A., & Pallier, C. (1993). Understanding compressed
sentences: the role of rhythm and meaning. Annals of the New
York Academy of Science, 682, 272-282.
Moon, C., Cooper, R. P., & Fifer, W. P. (1993). Two-day-olds
prefer their native language. Infant Behavior and Development,
16, 495-500.
Morgan, J. L., & Demuth, K. (1996). Signal to Syntax: Bootstrapping from speech to grammar in early acquisition. Mahwah NJ:
Lawrence Erlbaum Associates.
Morton, J., Marcus, S., & Frankish, C. (1976). Perceptual centers
(P-centers). Psychological Review, 83(5), 405-408.
Nazzi, T. (1997). Du rythme dans l’acquisition et le traitement
de la parole. doctoral dissertation, Ecole des Hautes Etudes en
Sciences Sociales.
Nazzi, T., Bertoncini, J., & Mehler, J. (1998). Language discrimination by newborns: towards an understanding of the role of
rhythm. Journal of Experimental Psychology: Human Perception and Performance, 24(3), 756-766.
Nespor, M., Guasti, M. T., & Christophe, A. (1996). Selecting word
order: The rhythmic activation principle. In U. Kleinhenz (Ed.),
Interfaces in Phonology (p. 1-26). Berlin: Akademie Verlag.
Pallier, C., Sebastian-Galles, N., Dupoux, E., Christophe, A., &
Mehler, J. (1998). Perceptual adjustment to time-compressed
speech: a cross-linguistic study. Memory & Cognition, 26, 844851.
Pijper, J. R. de. (1983). Modelling British English intonation. Dordrecht - Holland: Foris.
Pike, K. L. (1945). The intonation of American English. Ann Arbor,
Michigan: University of Michigan Press.
Pinker, S. (1984). Language learnability and language development. Cambridge: Harvard University Press.
Prince, A., & Smolensky, P. (1993). Optimality theory: Constraint
interaction in generative grammar (Tech. Rep. No. TR-2). Rutgers University.
Ramus, F., Dupoux, E., Zangl, R., & Mehler, J. (submitted). An
empirical study of the perception of language rhythm.
Ramus, F., Hauser, M., Miller, C., Morris, D., & Mehler, J. (2000).
Language discrimination by human newborns and by cotton-top
tamarin monkeys. Science, 288(5464), 349-351.
Ramus, F., & Mehler, J. (1999). Language identification with
suprasegmental cues: A study based on speech resynthesis.
Journal of the Acoustical Society of America, 105(1), 512-521.
Ramus, F., Nespor, M., & Mehler, J. (1999). Correlates of linguistic
rhythm in the speech signal. Cognition, 73(3), 265-292.
Roach, P. (1982). On the distinction between "stress-timed" and
"syllable-timed" languages. In D. Crystal (Ed.), Linguistic controversies. London: Edward Arnold.
14
FRANCK RAMUS
Sebastián-Gallés, N., Dupoux, E., & Costa, A. (2000). Adaptation
to time-compressed speech: Phonological determinants. Perception & Psychophysics, 4, 834-842.
Shi, R., Morgan, J. L., & Allopenna, P. (1998). Phonological and
acoustic bases for earliest grammatical category assignment: a
cross-linguistic perspective. Journal of Child Language, 25,
169-201.
Siqueland, E. R., & DeLucia, C. A. (1969). Visual reinforcement
of nonnutritive sucking in human infants. Science, 165(898),
1144-6.